Ray infra support in Gym

[sdevare-nv]

Use cases, pain points, and background
Gym should be able to use the Ray cluster to parallelize CPU-intensive tasks across multi nodes. This is necessary to support faster training and inference for environments like Comp coding or SWE agents.

Description:
Ray infra support in Gym

Design:

• Configure ray in gym to use the ray cluster deployed by NeMo-RL
• Refactor comp coding from using subprocess for verification to use ray.remote calls
• Refactor comp coding to use ray.remote calls over multiple GPU nodes
• Documentation on how to use ray in resource servers.

Out of scope:

• Configure detected CPU nodes for gym task. This will be part of NeMo-RL and will be worked on later

Acceptance Criteria:

• Speed-up in training step time for Comp coding environment